Flagging tasks likely to run late before they do, based on the owner's current load, the task's own history and upstream status.
Binary classification over task and owner features, trained on historical on-time versus late outcomes
scikit-learn RandomForestClassifier; simple historical-lateness rule as the baseline